sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/managers/cache_controller.py

@@ -18,7 +18,7 @@ import math
 import threading
 import time
 from queue import Empty, Full, PriorityQueue, Queue
-from typing import TYPE_CHECKING, List, Optional
+from typing import TYPE_CHECKING, List, NamedTuple, Optional, Set, Tuple
 
 import torch
 
@@ -33,6 +33,7 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
 )
 from sglang.srt.layers.dp_attention import (
+    get_attention_dp_rank,
     get_attention_tp_rank,
     get_attention_tp_size,
     is_dp_attention_enabled,
@@ -42,39 +43,53 @@ from sglang.srt.mem_cache.memory_pool import MHATokenToKVPool, MLATokenToKVPool
 logger = logging.getLogger(__name__)
 
 
+class LayerLoadingEvent:
+    def __init__(self, num_layers: int):
+        self._num_layers = num_layers
+        self.load_events = [torch.cuda.Event() for _ in range(num_layers)]
+        self.start_event = torch.cuda.Event()  # start event on controller stream
+
+    def complete(self, layer_index: int):
+        assert 0 <= layer_index < self._num_layers
+        self.load_events[layer_index].record()
+
+    def wait(self, layer_index: int):
+        torch.cuda.current_stream().wait_event(self.load_events[layer_index])
+
+    @property
+    def finish_event(self):
+        return self.load_events[-1]
+
+
 class LayerDoneCounter:
-    def __init__(self, num_layers):
+    def __init__(self, num_layers: int):
         self.num_layers = num_layers
         # extra producer and consumer counters for overlap mode
         self.num_counters = 3
-        self.counters = [num_layers] * self.num_counters
-        self.conditions = [threading.Condition() for _ in range(self.num_counters)]
-        self.producer_index = 0
-        self.consumer_index = 0
-
-    def next_producer(self):
-        return (self.producer_index + 1) % self.num_counters
+        self.events = [LayerLoadingEvent(num_layers) for _ in range(self.num_counters)]
+        self.producer_index = -1
+        self.consumer_index = -1
 
     def update_producer(self):
-        self.producer_index = self.next_producer()
+        self.producer_index = (self.producer_index + 1) % self.num_counters
+        assert self.events[
+            self.producer_index
+        ].finish_event.query(), (
+            "Producer finish event should be ready before being reused."
+        )
         return self.producer_index
 
-    def set_consumer(self, index):
+    def set_consumer(self, index: int):
         self.consumer_index = index
 
-    def increment(self):
-        with self.conditions[self.producer_index]:
-            self.counters[self.producer_index] += 1
-            self.conditions[self.producer_index].notify_all()
-
-    def wait_until(self, threshold):
-        with self.conditions[self.consumer_index]:
-            while self.counters[self.consumer_index] <= threshold:
-                self.conditions[self.consumer_index].wait()
+    def wait_until(self, threshold: int):
+        if self.consumer_index < 0:
+            return
+        self.events[self.consumer_index].wait(threshold)
 
     def reset(self):
-        with self.conditions[self.producer_index]:
-            self.counters[self.producer_index] = 0
+        self.producer_index = -1
+        self.consumer_index = -1
 
 
 class CacheOperation:
@@ -98,36 +113,30 @@ class CacheOperation:
         # default priority is the order of creation
         self.priority = priority if priority is not None else self.id
 
-    def merge(self, other: "CacheOperation") -> None:
-        # multiple operations can be merged into a single operation for batch processing
-        self.host_indices = torch.cat([self.host_indices, other.host_indices])
-        self.device_indices = torch.cat([self.device_indices, other.device_indices])
-        self.priority = min(self.priority, other.priority)
-        self.node_ids.extend(other.node_ids)
-
-    def split(self, factor) -> List["CacheOperation"]:
-        # split an operation into smaller operations to reduce the size of intermediate buffers
-        if factor <= 1:
-            return [self]
-
-        chunk_size = math.ceil(len(self.host_indices) / factor)
-        split_ops = []
-        for i in range(0, len(self.host_indices), chunk_size):
-            split_ops.append(
-                CacheOperation(
-                    host_indices=self.host_indices[i : i + chunk_size],
-                    device_indices=self.device_indices[i : i + chunk_size],
-                    node_id=0,
-                )
-            )
-        # Inherit the node_ids on the final chunk
-        if split_ops:
-            split_ops[-1].node_ids = self.node_ids
+    @staticmethod
+    def merge_ops(ops: List[CacheOperation]) -> CacheOperation:
+        assert len(ops) > 0
+        if len(ops) == 1:
+            return ops[0]
+
+        host_indices = torch.cat([op.host_indices for op in ops])
+        device_indices = torch.cat([op.device_indices for op in ops])
+        node_ids = []
+        priority = min(op.priority for op in ops)
+        for op in ops:
+            node_ids.extend(op.node_ids)
+        merged_op = CacheOperation(host_indices, device_indices, -1, priority)
+        merged_op.node_ids = node_ids
+        return merged_op
+
+    def __lt__(self, other: CacheOperation):
+        return self.priority < other.priority
 
-        return split_ops
 
-    def __lt__(self, other: "CacheOperation"):
-        return self.priority < other.priority
+class HiCacheAck(NamedTuple):
+    start_event: torch.cuda.Event
+    finish_event: torch.cuda.Event
+    node_ids: List[int]
 
 
 class TransferBuffer:
@@ -206,26 +215,25 @@ class PrefetchOperation(StorageOperation):
     ):
         self.request_id = request_id
 
-        self._done_flag = False
         self._lock = threading.Lock()
-
+        self._terminated_flag = False
         self.start_time = time.monotonic()
 
         super().__init__(host_indices, token_ids, last_hash)
 
     def increment(self, num_tokens: int):
         with self._lock:
-            if self._done_flag:
+            if self._terminated_flag:
                 return False
             self.completed_tokens += num_tokens
             return True
 
-    def mark_done(self):
+    def mark_terminate(self):
         with self._lock:
-            self._done_flag = True
+            self._terminated_flag = True
 
-    def is_done(self) -> bool:
-        return self._done_flag
+    def is_terminated(self) -> bool:
+        return self._terminated_flag
 
 
 class HiCacheController:
@@ -236,7 +244,7 @@ class HiCacheController:
         mem_pool_host: HostKVCache,
         page_size: int,
         tp_group: torch.distributed.ProcessGroup,
-        load_cache_event: threading.Event = None,
+        load_cache_event: threading.Event,
         write_policy: str = "write_through_selective",
         io_backend: str = "",
         storage_backend: Optional[str] = None,
@@ -340,8 +348,9 @@ class HiCacheController:
                 self.page_set_func = self._3fs_zero_copy_page_set
                 self.batch_exists_func = self._3fs_zero_copy_batch_exists
 
-        self.load_cache_event = load_cache_event
-        self.layer_done_counter = LayerDoneCounter(self.mem_pool_device.layer_num)
+        self.device = self.mem_pool_device.device
+        self.layer_num = self.mem_pool_device.layer_num
+        self.layer_done_counter = LayerDoneCounter(self.layer_num)
         self.mem_pool_device.register_layer_transfer_counter(self.layer_done_counter)
 
         if write_policy not in [
@@ -351,11 +360,11 @@ class HiCacheController:
         ]:
             raise ValueError(f"Invalid write policy: {write_policy}")
 
-        self.write_queue = PriorityQueue()
-        self.load_queue = PriorityQueue()
-
-        self.ack_write_queue = Queue()
-        self.ack_load_queue = Queue()
+        # self.write_queue = PriorityQueue[CacheOperation]()
+        self.load_queue: List[CacheOperation] = []
+        self.write_queue: List[CacheOperation] = []
+        self.ack_load_queue: List[HiCacheAck] = []
+        self.ack_write_queue: List[HiCacheAck] = []
 
         self.stop_event = threading.Event()
         self.write_buffer = TransferBuffer(self.stop_event)
@@ -366,16 +375,6 @@ class HiCacheController:
         self.write_stream = torch.cuda.Stream()
         self.load_stream = torch.cuda.Stream()
 
-        self.write_thread = threading.Thread(
-            target=self.write_thread_func_direct, daemon=True
-        )
-        self.load_thread = threading.Thread(
-            target=self.load_thread_func_layer_by_layer, daemon=True
-        )
-
-        self.write_thread.start()
-        self.load_thread.start()
-
         if self.enable_storage:
             self.prefetch_thread = threading.Thread(
                 target=self.prefetch_thread_func, daemon=True
@@ -402,9 +401,11 @@ class HiCacheController:
         if is_dp_attention_enabled():
             self.tp_rank = get_attention_tp_rank()
             self.tp_size = get_attention_tp_size()
+            self.dp_rank = get_attention_dp_rank()
         else:
             self.tp_rank = get_tensor_model_parallel_rank()
             self.tp_size = get_tensor_model_parallel_world_size()
+            self.dp_rank = 0
 
         # Currently, AscendMLAPagedTokenToKVPool is the subclass of MLATokenToKVPool.
         is_mla_backend = isinstance(self.mem_pool_device, MLATokenToKVPool)
@@ -430,15 +431,13 @@ class HiCacheController:
 
     def reset(self):
         self.stop_event.set()
-        self.write_thread.join()
-        self.load_thread.join()
 
-        self.write_queue.queue.clear()
-        self.load_queue.queue.clear()
+        self.write_queue.clear()
+        self.load_queue.clear()
         self.write_buffer.clear()
         self.load_buffer.clear()
-        self.ack_write_queue.queue.clear()
-        self.ack_load_queue.queue.clear()
+        self.ack_write_queue.clear()
+        self.ack_load_queue.clear()
         if self.enable_storage:
             self.prefetch_thread.join()
             self.backup_thread.join()
@@ -447,15 +446,7 @@ class HiCacheController:
             self.prefetch_revoke_queue.queue.clear()
             self.ack_backup_queue.queue.clear()
 
-        self.write_thread = threading.Thread(
-            target=self.write_thread_func_direct, daemon=True
-        )
-        self.load_thread = threading.Thread(
-            target=self.load_thread_func_layer_by_layer, daemon=True
-        )
         self.stop_event.clear()
-        self.write_thread.start()
-        self.load_thread.start()
 
         if self.enable_storage:
             self.prefetch_thread = threading.Thread(
@@ -471,7 +462,7 @@ class HiCacheController:
         self,
         device_indices: torch.Tensor,
         priority: Optional[int] = None,
-        node_id: int = 0,
+        node_id: int = -1,
     ) -> Optional[torch.Tensor]:
         """
         Back up KV caches from device memory to host memory.
@@ -480,17 +471,46 @@ class HiCacheController:
         if host_indices is None:
             return None
         self.mem_pool_host.protect_write(host_indices)
-        torch.cuda.current_stream().synchronize()
-        self.write_queue.put(
+        self.write_queue.append(
             CacheOperation(host_indices, device_indices, node_id, priority)
         )
+        self.start_writing()
         return host_indices
 
+    def start_writing(self) -> None:
+        if len(self.write_queue) == 0:
+            return
+
+        op = CacheOperation.merge_ops(self.write_queue)
+        host_indices, device_indices = self.move_indices(op)
+        self.write_queue.clear()
+
+        start_event = torch.cuda.Event()
+        finish_event = torch.cuda.Event()
+
+        start_event.record()
+        with torch.cuda.stream(self.write_stream):
+            start_event.wait(self.write_stream)
+            self.mem_pool_host.backup_from_device_all_layer(
+                self.mem_pool_device, host_indices, device_indices, self.io_backend
+            )
+            self.mem_pool_host.complete_io(op.host_indices)
+            finish_event.record()
+            # NOTE: We must save the host indices and device indices here,
+            # this is because we need to guarantee that these tensors are
+            # still alive when the write stream is executing.
+            if host_indices.is_cuda:
+                host_indices.record_stream(self.write_stream)
+            if device_indices.is_cuda:
+                device_indices.record_stream(self.write_stream)
+
+        self.ack_write_queue.append(HiCacheAck(start_event, finish_event, op.node_ids))
+
     def load(
         self,
         host_indices: torch.Tensor,
         priority: Optional[int] = None,
-        node_id: int = 0,
+        node_id: int = -1,
     ) -> Optional[torch.Tensor]:
         """
         Load KV caches from host memory to device memory.
@@ -499,76 +519,42 @@ class HiCacheController:
         if device_indices is None:
             return None
         self.mem_pool_host.protect_load(host_indices)
-        # to ensure the device indices are ready before accessed by another CUDA stream
-        torch.cuda.current_stream().synchronize()
-        self.load_queue.put(
+        self.load_queue.append(
             CacheOperation(host_indices, device_indices, node_id, priority)
         )
         return device_indices
 
-    def move_indices(self, host_indices, device_indices):
+    def move_indices(self, op: CacheOperation):
+        host_indices, device_indices = op.host_indices, op.device_indices
         # move indices to GPU if using kernels, to host if using direct indexing
         if self.io_backend == "kernel":
-            return host_indices.to(self.mem_pool_device.device), device_indices
+            if not host_indices.is_cuda:
+                host_indices = host_indices.to(self.device, non_blocking=True)
+            return host_indices, device_indices
         elif self.io_backend == "direct":
-            device_indices = device_indices.cpu()
-            host_indices, idx = host_indices.sort()
-            return host_indices, device_indices.index_select(0, idx)
+            if self.mem_pool_host.layout == "layer_first":
+                device_indices = device_indices.cpu()
+                host_indices, idx = host_indices.sort()
+                return host_indices, device_indices.index_select(0, idx)
+            elif self.mem_pool_host.layout == "page_first_direct":
+                return host_indices, device_indices.cpu()
         else:
             raise ValueError(f"Unsupported io backend")
 
-    def write_thread_func_direct(self):
-        """
-        Directly write through KV caches to host memory without buffering.
-        """
-        torch.cuda.set_stream(self.write_stream)
-        while not self.stop_event.is_set():
-            try:
-                operation = self.write_queue.get(block=True, timeout=1)
-                host_indices, device_indices = self.move_indices(
-                    operation.host_indices, operation.device_indices
-                )
-                self.mem_pool_host.backup_from_device_all_layer(
-                    self.mem_pool_device, host_indices, device_indices, self.io_backend
-                )
-                self.write_stream.synchronize()
-                self.mem_pool_host.complete_io(operation.host_indices)
-                for node_id in operation.node_ids:
-                    if node_id != 0:
-                        self.ack_write_queue.put(node_id)
-            except Empty:
-                continue
-            except Exception as e:
-                logger.error(e)
+    def start_loading(self) -> int:
+        if len(self.load_queue) == 0:
+            return -1
 
-    def load_thread_func_layer_by_layer(self):
-        """
-        Load KV caches from host memory to device memory layer by layer.
-        """
-        torch.cuda.set_stream(self.load_stream)
-        while not self.stop_event.is_set():
-            self.load_cache_event.wait(timeout=1)
-            if not self.load_cache_event.is_set():
-                continue
-            self.load_cache_event.clear()
-            self.layer_done_counter.update_producer()
-
-            batch_operation = None
-            while self.load_queue.qsize() > 0:
-                op = self.load_queue.get(block=True)
-                if batch_operation is None:
-                    batch_operation = op
-                else:
-                    batch_operation.merge(op)
-            if batch_operation is None:
-                continue
+        producer_id = self.layer_done_counter.update_producer()
+        op = CacheOperation.merge_ops(self.load_queue)
+        host_indices, device_indices = self.move_indices(op)
+        self.load_queue.clear()
+        producer_event = self.layer_done_counter.events[producer_id]
+        producer_event.start_event.record()
 
-            # start layer-wise KV cache transfer from CPU to GPU
-            self.layer_done_counter.reset()
-            host_indices, device_indices = self.move_indices(
-                batch_operation.host_indices, batch_operation.device_indices
-            )
-            for i in range(self.mem_pool_host.layer_num):
+        with torch.cuda.stream(self.load_stream):
+            producer_event.start_event.wait(self.load_stream)
+            for i in range(self.layer_num):
                 self.mem_pool_host.load_to_device_per_layer(
                     self.mem_pool_device,
                     host_indices,
@@ -576,13 +562,24 @@ class HiCacheController:
                     i,
                     self.io_backend,
                 )
-                self.load_stream.synchronize()
-                self.layer_done_counter.increment()
-
-            self.mem_pool_host.complete_io(batch_operation.host_indices)
-            for node_id in batch_operation.node_ids:
-                if node_id != 0:
-                    self.ack_load_queue.put(node_id)
+                producer_event.complete(i)
+            self.mem_pool_host.complete_io(op.host_indices)
+            # NOTE: We must save the host indices and device indices here,
+            # this is because we need to guarantee that these tensors are
+            # still alive when the load stream is executing.
+            if host_indices.is_cuda:
+                host_indices.record_stream(self.load_stream)
+            if device_indices.is_cuda:
+                device_indices.record_stream(self.load_stream)
+
+        self.ack_load_queue.append(
+            HiCacheAck(
+                start_event=producer_event.start_event,
+                finish_event=producer_event.finish_event,
+                node_ids=op.node_ids,
+            )
+        )
+        return producer_id
 
     def evict_device(
         self, device_indices: torch.Tensor, host_indices: torch.Tensor
@@ -625,7 +622,7 @@ class HiCacheController:
         return operation
 
     def terminate_prefetch(self, operation):
-        operation.mark_done()
+        operation.mark_terminate()
         return operation.completed_tokens, operation.hash_value
 
     def append_host_mem_release(self, host_indices: torch.Tensor):
@@ -706,6 +703,7 @@ class HiCacheController:
                 operation.completed_tokens
                 != prev_completed_tokens + len(batch_hashes) * self.page_size
             ):
+                operation.mark_terminate()
                 break  # Some operations fail or operation terminated by controller
         # release pre-allocated memory
         self.append_host_mem_release(
@@ -885,7 +883,7 @@ class HiCacheController:
 
                 if not self.backup_skip:
                     self._page_backup(operation)
-                self.ack_backup_queue.put(operation.id)
+                self.ack_backup_queue.put(operation)
 
             except Empty:
                 continue